Research Engineer/Scientist

Own LLM mid-training and post-training research, including continued pretraining, SFT, preference optimization, and RL; make data-mixture and experimental decisions and determine how training changes affect downstream agent behavior.
Research and prototype novel agentic architectures and algorithms across planning, reasoning, memory, skills, tool use, retrieval, and multi-agent collaboration, advancing beyond existing approaches where appropriate.
Design and build research harnesses and experimentation methodologies that enable systematic experimentation, trajectory analysis, reproducibility, and rigorous comparison across models, checkpoints, and agent architectures.
Define evaluation methodologies and develop novel benchmarks for measuring agent reasoning, planning, tool use, reliability, factuality, and safety; establish rigorous approaches for LLM-as-a-Judge, trajectory-based, and human evaluation.
Identify systematic model and agent failure patterns, determine their root causes, and translate those insights into new research directions or improvements in training data, architecture, context, or evaluation.
Independently identify research problems, formulate novel hypotheses, and drive projects from research idea to validated prototype and measurable impact, collaborating with engineering to transition successful approaches into production and communicating results through publications, patents, or open-source work.
3+ years of experience in machine learning, deep learning, AI research, or a related field, with demonstrated applied research experience and a track record of independently driving research projects.
Hands-on experience with LLM training and post-training, including one or more of continued pretraining, SFT, preference optimization, or RL; experience making training-data decisions and understanding training dynamics and failure modes at scale.
Strong Python and advanced PyTorch expertise, with experience modifying models, training pipelines, or research infrastructure to support novel experimentation.
Strong practical depth in agentic AI and context engineering, including planning, reasoning, memory, skills, tool use, retrieval, long-context processing, and knowledge grounding.
Experience designing evaluation methodologies, not just running evaluations, including benchmark design, trajectory-based evaluation, LLM-as-a-Judge, human evaluation, and the ability to determine which metrics and methodologies are appropriate for a research question.
Demonstrated research ownership and impact, with the ability to identify important research questions, develop novel hypotheses, conduct rigorous experiments, and communicate findings and implications effectively to technical researchers, engineers, and executive stakeholders

Employment type
Full-time

Work arrangement
No